Risky Driving Behavior Recognition Based On Artificial Intelligence

With the increase of car ownership,road traffic safety has become a key issue affecting social and economic development.The occurrence of road traffic accidents poses a great threat to the safety of people’s lives and property.The frequency and severity of road traffic accidents are relatively high on the expressways,leading to poor traffic safety condition.The risky driving behavior of drivers is an important cause of traffic accidents.How to identify risky driving behavior quickly and accurately has become one of the key issues of road traffic safety.The paper focuses on the most common risky behaviors,risky car-following behavior and risky lane-changing behavior.With trajectory data and artificial intelligent algorithms,we respectively established risky car-following behavior and risky lane-changing behavior recognition models.The research can recognize the risky driving behaviors rapidly and accurately,decrease the road traffic safety,and provide scientific basis and decision support for road traffic safety,the formulation of traffic safety policies,the design of driving assistance system and the development of automatic driving.Firstly,based on the vehicle trajectory data of expressway video,the paper extracts car-following samples and lane-changing samples as the research object.By matching the identification of car-following vehicles and lane-changing vehicles with surrounding vehicles,the relative speed and relative distance between the target vehicle and surrounding vehicles are determined.Then the driving behavior dataset is established with trajectory data,and the distribution features of trajectory data is analyzed.At the same time,the samples are labeled based on traffic context,determining the vehicle type and traffic flow type of the samples,which provides the data for the multi-feature integration intelligent recognition model.Secondly,the risk analysis and expert evaluation method are combined to label the risk of car-following behavior and lane-changing behavior.As for the car-following behavior,three risk parameters,including headway,collision margin and deceleration to avoid crashes,are used for risk evaluation.As for the lane-changing behavior,the lane change risk parameters are used to evaluate the risk.Then the t-distributed Stochastic Neighbor Embedding(t-SNE)algorithm is used to visualize the risk of carfollowing and lane-changing behaviors.Based on the above research,it can provide reliable label results for risky driving behavior.In addition,the discrete Fourier features,discrete wavelet features and statistical features of car-following and lane-changing behavior are extracted as key feature parameters to provide input variables for risky driving behavior recognition model.Based on the label results of car-following behavior and the key features of trajectory variables,the risky driving behavior recognition model is established.We adopt six intelligent algorithms,including XGBoost,Ada Boost,LGBM,random forest,multilayer perceptron,and support vector machine to establish the model.Then the influence of traffic contexts,including vehicle type and traffic flow type,on carfollowing behavior is analyzed by comparing the trajectory variables distribution under different traffic contexts.The vehicle type recognition model and traffic flow recognition model are respectively established based on the key features of trajectory variables.Then the recognition results of the two models are integrated into the key features of car-following trajectory variables as the new inputs of the risky carfollowing recognition model.By comparing different models,the performance of twofeatures integration model achieves best,of which the recognition accuracy arrives96.25%.Based on the label results of lane-changing behavior and the key features of trajectory variables,the risky lane-changing recognition model is proposed.Then we further analyze the influence of traffic contexts on lane-changing behavior,including vehicle type,traffic flow type,and lane-changing type.The left lane-changing and right lane-changing behaviors can lead to different lane change risk.Therefore,the lanechanging behavior recognition model is established based on trajectory variables to classify the lane-changing type.The vehicle type and traffic flow type recognition model is the same as those of car-following behavior.The three traffic contexts are integrated into the key features of trajectory variables to establish multi-features integration model to recognize the risky lane-changing behavior.By comparing different models,it indicates that the three-features integration model achieves the best performance,of which the accuracy is 85.03%.Finally,based on the risky car-following behavior and risky lane-changing behavior recognition model with multi-feature integration,this paper further analyzes the non-linear relationships between trajectory features and driving behavior risk.The interpretable machine learning model was applied to explain the risky driving behavior intelligent recognition model.Firstly,the importance of relevant independent variables in the recognition model is analyzed by using the permutation feature importance model.The factors that have significant impacts on the recognition results of risky driving behavior are identified.Then the non-linear influence of significant factors on risky driving behavior is explored with accumulated local effects(ALE)model.The interaction effect of two factors is also investigated with ALE model.Based on the above research,we can explore the formation mechanism of risky driving behavior and provide theoretical support for the risky driving behavior prevention.This paper establishes the risky car-following and risky lane-changing recognition model with multi-features integration,which improves the recognition accuracy.The formation mechanism of risky driving behavior is analyzed.Rapid and accurate recognition of risky driving behavior can provide theoretical support for the development of driving assistance system,traffic safety decision-making and automatic driving technology,so as to improve the level of road traffic safety and operation efficiency.